The examples are Julia scripts that you can run. Some as similar to the tutorials, others explore additional aspects of HTBoost.
A good way to familiarize yourself with HTBoost and compare (its performance to LigthGBM) is to study and run the following examples:
- Basic use (main options, cv, savings and loading results, variable importance and more post-estimation analysis)
- Logistic (binary classification)
- Global Equity Panel (time series and panels/longitudinal data, with various options for cv)
- Categoricals (how HTBoost handles categorical features)
- Missing data (HTBoost excels at handling missing data)
- Speeding up with large n (strategies to reduce computing time for large n)
The other examples explore more specific aspects of HTBoost:
Understanding hybrid trees
- Hybrid trees (how HTBoost can escape local minima of smoothtrees)
- Projection pursuit regression (an example where adding a single index model to each tree (the default in HTBoost) improves forecasting)
Other distributions (loss functions)
- Multiclass (multiclass classification)
- Zero inflated y (y≥0, continuous except for positive mass at 0)
- GammaPoisson (aka negative binomial for count data)
- Huber and t unbiased (outlier robust losses in HTBoost and lightGBM)
- t distribution (the recommended robust loss in HTBoost)
- Gamma distribution (discusses options if min(y)>0)
Others
- Offset (exposure) (how to add an offset, common in e.g. insurance, biology ...)
- Sparsity penalization (how HTBoost improves forecasting by feature selection when p is large)
- Speedups with sparsevs (how HTBoost speeds up feature selection when p is large)